A music detector, a design ranking built from human preferences and a one-off 3D scene ask different questions about the quality and origin of AI output.
Artificial Intelligence··Morning
A dispute over the origin of a song
Fenix Flexin's song 'Rubberz,' released on 5 June, has been flagged as likely AI-generated by the new detector inside the music app Treblo. The track is No. 58 on the Billboard Hot 100 and its video has 7 million views. None of the well-known AI music detectors had flagged it through the summer. Independent producer Medasin argued that the app dropped the Sonauto name two days before the song's release, that a freeze-frame from an older Instagram story showed a filename ending in 'Sonauto.mp3,' and that an '80s synthwave' tag could produce similar lyrics. A WIRED reporter repeating the attempt obtained near-identical couplets. Treblo's new detector now flags 'Rubberz' and some songs by Tyga as likely AI-generated. Fenix Flexin and producer Purps on the Beat deny using AI. The artist said he freestyled the lyrics and posted what he described as the original project file; producers commenting on it disagreed about what the file established. The detector flag, the filename, the similar-lyric attempt and the artist's denial therefore sit in one dispute while supplying different kinds of information.[1]
Human preference in design outputs
Design Arena uses a different evaluation arrangement: it shows users two visual outputs and asks which one they prefer. AI models that generate websites, images and similar formats are ranked from those head-to-head choices. The resulting preference data is used by laboratories to improve their models. Intelligence, the company behind the platform, announced a $7.9 million seed round led by Index Ventures; Conviction, A* and Valkyrie also participated. Co-founder Grace Li says this kind of feedback was the missing bottleneck for many models trying to improve in design. The report does not name the models evaluated on the platform. This arrangement measures something different from a Treblo detector seeking the origin of a song. Design Arena's data shows which of two presented outputs people choose and turns those choices into a relative ranking. The ranking converts preference within a supplied pair into a structured feedback stream. The company's new capital supports the expansion of preference collection and model ranking as a separate product and data layer. The central development here is that human taste is being organised into data that can travel back to model developers, while the company operating that process has secured new funding.[2]
From a single demonstration to a capability signal
Andrej Karpathy's experiment is neither a detector nor a human-preference ranking, but an informal capability demonstration built around one task. Karpathy gave Claude Opus 5 a paragraph from 'The Lord of the Rings' and asked it to build a browser-based 3D scene with Three.js. According to the report, the model produced 5,500 lines of code in about two hours at a cost of roughly $10. The result was posted on X and at karpathy.ai/lotr-movie/. Karpathy places the work in a line of earlier informal tests, including GPT-4 drawing a unicorn in TikZ in 2023 and an SVG pelican on a bicycle. There is no reported baseline, repeatability condition or scoring method; the account describes it as a one-off demonstration rather than a benchmark. Put together, the three developments show that the word evaluation can carry different questions. Treblo's flag estimates the possible origin of an output. Design Arena gathers people's choices between two alternatives. Karpathy's scene demonstrates what can be produced from a particular prompt. An estimate of provenance, a ranking of relative taste and a demonstration of possibility are not interchangeable. They provide three separate signals about AI output, each tied to a different procedure and a different question.[1], [2], [3]
Related columns
For more information on this topic, you can read the related columns.
